Recurrence and Plasticity in Evolved Neural Controllers

Among the more important applications of evolutionary neurocontrollers is the development of systems that are able to dynamically adapt to a changing environment. While traditional approaches to control system design demand that the developer attempt to foresee all possible situations within which the controller may operate, neuroevolutionary approaches can facilitate the design of systems that are capable of operating in unforeseen circumstances. This paper examines two methods that have been used to provide for this adaptivity. The first method is the use of recurrent neural networks that have fixed connection weights. The second develops neurocontrollers with plastic synapses, thus allowing for the adaptation of the connection weights. Previous experimental results have shown that while both approaches can facilitate adaptive behavior, neural plasticity does not necessarily confer the expected benefits. In experiments using the NeuroEvolution of Augmenting Topologies (NEAT) method, Stanley et al. (2003) discovered that in simple cases, recurrence was sufficient in solving at least some control problems. I examined whether or not these initial results continue to scale upwards into more complex problem spaces. This was done through a series of experiments ranging from controlling a simplified cannon shot to attempting to evolve neural flight controllers capable of flying different airplanes through a series of waypoints. The results of these experiments indicate that the NEAT algorithm itself is unable to scale efficiently to some larger problem spaces.  M.S